Emitted each time a swarm child resolves, so a client can render a live
“N/total done” progress tick between AgentSwarmStarted and
AgentSwarmCompleted rather than only the final result.
A running tally of swarm children as they resolve, used for live progress.
resolved is completed + failed + aborted; clients can render
resolved/total as a progress bar without tracking individual children.
Emitted when the agent_swarm tool begins fanning out children, so any
app-server/SDK/TUI client can observe the swarm as a whole. Per-child
progress flows through the existing Subagent* trace events.
Emitted when ultra mode is toggled on a runtime/thread. Ultra mode enables
proactive multi-agent delegation for any model (not only Codex Sol/Terra
Ultra effort).
How ultra mode was entered. manual is a persistent /ultra on toggle;
task is a one-shot /ultra <prompt>. Only manual stays active across
turns. Selecting Codex Sol/Terra Ultra reasoning still enables proactive
multi-agent for that model without flipping this mode flag.
Canonical swarm-mode reminder injected into a turn’s developer instructions
(server-side) while agent-swarm mode is active, so the model reaches for the
agent_swarm fanout tool. Shared by the runtime injection and the TUI label
so the wording stays in one place.
Default quiet window after which the swarm recovers one unit of global
rate-limit capacity, in milliseconds (3 minutes). Any fresh rate limit
restarts the window, and recovery happens at most once per quiet window.
Default minimum interval between successive global rate-limit capacity
shrinks, in milliseconds. While the provider keeps rate-limiting, the swarm
shrinks its global capacity by one at most this often, so it throttles
smoothly instead of collapsing on the first burst of 429s.
Canonical model-facing swarm tool name. Single source of truth shared by
the tool registration (roder-ext-subagents) and the core turn loop’s
exclusivity enforcement (roder-core).
Canonical ultra-mode reminder for one-shot /ultra <prompt> turns. Persistent
ultra mode injects proactive multi-agent policy server-side instead of
prepending this to the user-visible transcript.
Sink the runtime supplies on a tool-execution context so the agent_swarm
tool can publish live progress snapshots (carrying the runtime’s thread/turn
ids and the bus emitter) without the tool depending on roder-core.
Detect whether a batch of tool-call names violates the agent_swarm
exclusivity rule. Returns None when the batch is valid: no agent_swarm
call, or exactly one agent_swarm call by itself.